zucchetti-llm-course

Lab 02: Prompt Engineering Workshop

Lecture: 3, Prompt Engineering Duration: 35 minutes Difficulty: Beginner

Objectives

Understand why naive prompts fail (hallucination, verbosity, wrong format)
Practice zero-shot, few-shot, and Chain-of-Thought prompting
Learn to force structured JSON output from LLMs
Compare 5 different prompt strategies on the same query
Experiment with Temperature and Top-P to understand generation parameters
Understand prompt injection attacks and how to defend against them

Prerequisites

Requirement	How to Check
Python 3.10+	`python --version`
Dependencies	`pip install openai rich python-dotenv`
LLM access	Google AI Studio API key in `.env` as `GOOGLE_API_KEY`

Get a free API key: Visit aistudio.google.com/apikey — takes 30 seconds.

Note: This lab does NOT require RAG or a vector database. Context is provided manually so you can focus 100% on prompting techniques. RAG is covered in the next lab (Lab 03).

Quick Start

cd lab_02_prompting
python lab_02_prompting.py

Files

File	Description
`lab_02_prompting.py`	Main script — 10 steps, 5 prompt strategies compared side-by-side
`lab_02_prompting.ipynb`	Jupyter notebook version
`solutions.py`	Reference implementation for the Chain-of-Thought TODO

Lab Steps

Step	Topic	What You’ll See
1	Hello World	API connection test
2	Zero-Shot	Verbose, unpredictable output (the “before”)
3	System Message	One-word answer, controlled output (the “after”)
4	Few-Shot	Format-perfect output from examples alone
5	Chain-of-Thought	Step-by-step reasoning shown explicitly
6	Structured JSON	Production-ready parseable output
7	Grand Comparison	All 5 strategies side-by-side
8	Temperature	Deterministic → creative → chaotic
8b	Top-P & max_tokens	Vocabulary pool and output length controls
9	Prompt Injection	Real attack attempts on a scoped bot
9b	Scope Guard Defense	Prompt chaining as a security mechanism
10	Production Prompt	System + Few-Shot + CoT + JSON combined

Tasks

Task 1: Run All Steps

Run the script end-to-end and observe how the same review produces different answers with each prompt strategy.

Task 2: Observe the Temperature Effect (Step 8)

Compare outputs at T=0.0, T=0.5, T=1.0, T=1.5. At what point does the output become unusable?

Task 3: Try a New Injection Attack (Step 9)

Modify one of the three attack strings and see if Gemini resists your custom attack.

Task 4: Implement Chain-of-Thought (Bonus)

Find the step_5_cot function and try writing your own CoT prompt before looking at solutions.py.

Key Concepts

Strategy	When to Use
Zero-Shot	Quick experiments, not for production
System Message	Always — the single highest-return improvement
Few-Shot	When you need consistent format or tone
Chain-of-Thought	Complex reasoning; large models only
Structured Output	Any time downstream code needs to parse the answer
Prompt Chaining	Complex multi-step tasks; security (scope guard)